HikmaAI DistilBERT PII NER

A fine-tuned DistilBERT model for Named Entity Recognition (NER) of Personally Identifiable Information (PII) in multilingual text. Trained on 6 languages (English, Italian, French, German, Spanish, Dutch). Detects 17 entity types across personal, financial, and contact information categories.

Model Details

Entity Types

Usage

Transformers

from transformers import pipeline

ner = pipeline(
    "ner",
    model="hikmaai-io/hikmaai-distilbert-pii",
    aggregation_strategy="simple"
)

results = ner("Contact John Smith at john@example.com or call +1-555-0123")
for entity in results:
    print(f"{entity['word']:20s} -> {entity['entity_group']:20s} ({entity['score']:.2f})")

ONNX Runtime

import onnxruntime as ort
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("hikmaai-io/hikmaai-distilbert-pii")
session = ort.InferenceSession("onnx/model_quantized.onnx")

inputs = tokenizer("Contact John Smith at john@example.com", return_tensors="np")
outputs = session.run(None, {
    "input_ids": inputs["input_ids"],
    "attention_mask": inputs["attention_mask"],
})
# outputs[0] shape: [1, seq_len, 35] (logits for 35 BIO labels)

ONNX Runtime (Go)

The model is designed for high-performance inference in Go applications via yalue/onnxruntime_go:

// 2 input tensors: input_ids, attention_mask
// Output: logits [1, seq_len, 35]
// Post-processing: argmax per token -> BIO label -> entity spans

Training

• Base model: distilbert-base-uncased
• Dataset: ai4privacy/pii-masking-400k (325K annotated examples, 6 languages: EN 21%, IT 20%, FR 20%, DE 20%, ES 10%, NL 10%)
• Checkpoints: 15,000 steps
• Quantization: INT8 dynamic quantization via ONNX Runtime (FP32 also available)

File Structure

├── config.json              # Model configuration
├── model.safetensors        # FP32 weights (PyTorch/safetensors)
├── tokenizer.json           # Fast tokenizer
├── tokenizer_config.json    # Tokenizer configuration
├── onnx/
│   ├── model.onnx           # FP32 ONNX
│   └── model_quantized.onnx # INT8 ONNX (recommended for inference)

Privacy Design

This model is designed for privacy-preserving PII detection in AI security gateways:

• Detection only: the model identifies entity spans and types. It does not store, transmit, or log the actual PII values
• Counts, not content: downstream systems receive entity type counts (e.g., {"EMAIL": 2, "PHONE": 1}), never the detected text
• Local inference: runs entirely on-device via ONNX Runtime, no external API calls
• Stateless: no request data is retained between inferences

Limitations

• Trained on 6 languages (EN, IT, FR, DE, ES, NL); performance on other languages is untested
• Maximum sequence length: 512 tokens (longer texts require chunking)
• Regex-based PII detection (emails, phone numbers, credit cards) may outperform NER for well-structured patterns; this model adds value for unstructured PII (names, addresses, free-text credentials)
• INT8 quantization may slightly reduce accuracy on edge cases compared to FP32

Citation

@misc{hikmaai-distilbert-pii,
  author = {HikmaAI},
  title = {HikmaAI DistilBERT PII NER: Fine-tuned model for PII detection in AI security gateways},
  year = {2026},
  url = {https://huggingface.co/hikmaai-io/hikmaai-distilbert-pii}
}